IS6335 Chap.3 Data Abstraction, Attribute Types and Task Targets
Data Abstraction, Attribute Types and Task Targets
Three dataset types, and only three
Answering the what question means placing your data in one of three families. The course names them tables, networks and spatial, and the small vocabulary is deliberate, because the idioms of week four onward are organised by exactly these three.
A table has items as rows, attributes as columns and a cell where the two meet, and a multidimensional table is keyed on more than one key. A network has items as nodes, also called vertices, joined by links, also called edges or connections, with trees as the special case that has no cycles.
Spatial data splits again: a continuous field is a grid of positions each carrying attributes, while geometry is shape given directly in space.
One line beside the spatial family repays attention. In visualization, as opposed to computer graphics, geometry is a design decision.
In a geometry dataset the shape is given and you are obliged to keep it; in a table you invent the positions and are therefore responsible for them. Knowing which situation you are in tells you whether position is available as a channel at all.
A further axis is named alongside the types, dataset availability, which is either static or dynamic, and a dynamic dataset rules out any design whose layout would jump when a new item arrives.
Attribute types decide which channels are legal
Within a dataset every attribute has a type, and the type decides what may carry it. The split is two stage.
An attribute is categorical, with no meaningful order, or ordered. An ordered attribute is ordinal, where rank is meaningful but arithmetic is not, or quantitative, where differences and ratios can be computed.
An ordered attribute also has a direction: sequential, running from low to high; diverging, with a meaningful midpoint and two directions away from it; or cyclic, returning to where it started.
Two failures follow directly from getting this wrong and both produce charts that render perfectly.
Diverging data drawn on a sequential ramp puts the meaningful zero at an arbitrary shade, so the reader loses the only boundary that matters. Cyclic data drawn on a straight axis puts December and January at opposite ends of the chart when they are one step apart. A third failure is quieter: postal codes, room numbers and survey response codes are stored as numbers and are categorical, and a tool will happily average them.
Before accepting a mean, ask whether the midpoint between two values of that attribute is itself a value of that attribute.
The task half: actions and targets
The why question is answered with a pair, an action and a target.
High-level actions divide into consuming, which covers discovering something new, presenting something already known and a third that the list simply labels enjoy, and producing, which covers annotating, recording and deriving. At the lower level, search is described in terms of what the reader already knows: is the target known or unknown, and is its location known or unknown.
Those two binary questions give four search situations, and a second dimension asks how much of the data matters, which is one item, some items or all of them. The lecture is explicit that these are independent choices that mix and match.
Under producing sits the most consequential instruction in the course. Do not simply draw what arrived.
Work out what ought to be shown, then build it by transforming the original dataset in a series of steps. Deriving is named as one of four principal strategies for coping with complexity, and the worked case is an eighteenth century balance-of-trade chart in which exports and imports are the original data and the trade balance, their difference, is derived.
Nothing was added to the dataset; a quantity already implied by it was made visible.
Targets, and the pair that conflict
Targets have their own vocabulary and it changes with the dataset type. For data in general the targets are trends, outliers and features. For attributes they are distribution, dependency, correlation and outliers. For networks they are topology and path.
For spatial data the target is shape or distribution.
The last two network targets are the clearest design conflict in the course. Seeing topology wants the whole graph on screen, which pushes nodes together and makes individual edges impossible to follow. Tracing a path wants edges separable, which pushes nodes apart and pushes most of the graph out of view.
One static view cannot serve both well, so the honest answer is two views or an interaction that moves between them. Recognising the conflict is worth more than resolving it badly, and the same reasoning is what makes the adjacency matrix the right answer two chapters later.
What this chapter covers
- 01
Tables, networks and spatial as the three dataset types
- 02
Items, attributes, cells, nodes, links, fields and geometry
- 03
Static against dynamic dataset availability
- 04
Categorical, ordinal and quantitative attributes
- 05
Sequential, diverging and cyclic ordering directions
- 06
Actions: consume and produce, search and query
- 07
Derive as a strategy for handling complexity
- 08
Targets, and the topology against path conflict
Type a real file and write its task abstraction
- 2Name the dataset type and say what one item is.
- 3Type all six attributes.
- 2Write the action and the target.
- 1Name one attribute that would mislead a tool.
Key terms
- Item
- One row of a table, or one node of a network. Deciding what an item is comes before everything else, because a count of items answers a different question from a sum over an attribute.
- Ordinal attribute
- An ordered attribute whose rank is meaningful but whose differences are not. A satisfaction scale can be ranked and sorted; the gap between two of its levels is not a measured quantity.
- Diverging attribute
- An ordered attribute with a meaningful midpoint and two directions away from it, such as a change against a target. Its colour scale must be centred on that midpoint rather than on the observed range.
- Continuous field
- A spatial dataset given as a grid of positions, each cell carrying attributes. Its positions are sampled, so resolution is inherited rather than chosen.
- Target
- What the reader is looking for: a trend, an outlier, a feature, a distribution, a dependency, a topology, a path or a shape. It is the second half of a task abstraction.
- Derived attribute
- A column created by transforming the ones supplied, such as a difference or a rate. The attribute a question is really about is frequently one that does not exist in the file yet.
Data Abstraction, Attribute Types and Task Targets FAQ
How do I decide whether my data is a table or a network?
By the question, not by the file. The same file of taxi trips is a table if you are asking how fares are distributed by hour, because items are trips and fare and hour are attributes; it is a network if you are asking which locations feed which, because nodes are locations and links are trips; and it is spatial if you are asking where demand concentrates.
One physical file can carry three possible data abstractions, and the task decides which one is in force. What you cannot do is leave the decision implicit, because each of the three brings a different set of idioms with it.
Is a mean of a satisfaction score wrong?
It is a convention rather than an error, and the safe habit is to say which you are relying on. The score is ordinal, so its rank is meaningful and its differences are not, which means a mean assumes that the gap between the first and second levels is the same as the gap between the fourth and fifth. That assumption is often reasonable and occasionally destroys the finding.
Reporting the distribution alongside the mean costs nothing and is what the outlier target asks for anyway, since a bimodal set of scores can produce a comfortable-looking average that describes nobody.
Assessment move
Build the habit of typing a file before opening any chart tool. Write one line per attribute with its type and its ordering direction, and mark the numeric ones that are actually categorical.
Ten minutes of that at the start of a lab exercise prevents the two most common failures in the weeks that follow, a summed identifier and a diverging attribute on a sequential scale, and it gives you most of the data source description the individual report is marked on.
Working through Data Abstraction, Attribute Types and Task Targets in IS6335? Sia is AskSia’s AI Information Technology tutor — ask any IS6335 Data Abstraction, Attribute Types and Task Targets question and get a clear, step-by-step explanation grounded in how IS6335 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
